Modeling the Background for Incremental Learning in Semantic Segmentation

Despite their effectiveness in a wide range of tasks, deep architectures\nsuffer from some important limitations. In particular, they are vulnerable to\ncatastrophic forgetting, i.e. they perform poorly when they are required to\nupdate their model as new classes are available but the original training set\nis not retained. This paper addresses this problem in the context of semantic\nsegmentation. Current strategies fail on this task because they do not consider\na peculiar aspect of semantic segmentation: since each training step provides\nannotation only for a subset of all possible classes, pixels of the background\nclass (i.e. pixels that do not belong to any other classes) exhibit a semantic\ndistribution shift. In this work we revisit classical incremental learning\nmethods, proposing a new distillation-based framework which explicitly accounts\nfor this shift. Furthermore, we introduce a novel strategy to initialize\nclassifier's parameters, thus preventing biased predictions toward the\nbackground class. We demonstrate the effectiveness of our approach with an\nextensive evaluation on the Pascal-VOC 2012 and ADE20K datasets, significantly\noutperforming state of the art incremental learning methods.\n